OrgaTuring: A Novel Vision-AI Assisted Organoid Mining Towards Accelerating Disease Discovery

Introduction: Organoid cultures are 3D in vitro tissue construct that emulates their corresponding in vivo organ. Organoids’ accurate mimicking nature has made them powerful in vitro models to study various aspects of a tissue. Organoids are generally grown in a 3D setup using either naturally derived or synthetic extracellular matrices. They are commonly studied by investigating their morphological features and growth characteristics. However, such a practice is very challenging due to the inherent imaging artifacts in organoid images. Recently, very few segmentation techniques have been introduced in the literature to perform localization and quantification of organoids. Unfortunately, no attempts have been made to reliably classify healthy and diseased organoids or to predict ailments in an organoid. In this work, we demonstrate OrgaTuring, an end-to-end vision AI approach that can locate, quantify, and classify human colon organoids with an accuracy of
90%. OrgaTuring can serve as a completely automated computational framework to investigate thousands of images with no expert intervention.


Method: OrgaTuring comprises (1) a novel convolutional neural network pipeline with explainability using PyTorch; and (2) a manually labeled human colon organoid image dataset. We have made the deep learning model, inference procedures, and image dataset publicly available along with a detailed manual for easy adoption.


Results: OrgaTuring, using a custom-built YOLOv5 vision AI model, can accurately count the number of organoid tissue in an image within milliseconds. Further, it utilizes a DenseNet201 pre-trained Imagenet architecture for healthy vs disease organoid classification (in particular Chron's diseased organoids). Results indicate a training accuracy of 87%, validation accuracy of 89%, and testing accuracy of 89%, with a ROC/AUC of 90%.


Discussion: Most computational tools and techniques used to study organoids focus only on quantification. Contrary to that, OrgaTuring, with its unique AI algorithms, goes beyond counting (or quantification). It can also locate, track, and classify organoids w.r.t various meaningful phenotypes. This makes OrgaTuring an all-encompassing AI-guided toolbox for rapid organoid discovery. Further, OrgaTuring real-time nature is a boon to system biologists to predict outcomes in milliseconds without relying on too many expert interventions. Our model bears enough potential to classify thousands of images obtained from different imaging techniques, parameters, and cohorts in real-time. As for the results go, OrgaTuring is able to classify healthy vs diseased organoids within milliseconds. At this point, we focus only on colon organoids. However, OrgaTuring bears the capacity to handle other types of organoids too.


Conclusion: Organoids replicate specific diseases' cellular and molecular characteristics, provide valuable insights into disease mechanisms, identify drug targets, and test potential therapeutic approaches. This accelerates the early stages of drug development by providing a more accurate representation of human disease in laboratory settings. In this context, OrgaTuring could be a catalyst in accelerating Organoid-based disease diagnosis, therapeutics, and drug discovery.